Logging formation division method and device based on hybrid deep learning and geological constraints

Abstract

The application discloses a well logging stratum division method and device based on mixed deep learning and geological constraints, and relates to the field of stratum identification and division. First, well logging data is preprocessed, and the well logging data comprises well logging curves and stratum marker point depths. Second, initial data is input into a pre-trained stratum division model for prediction, and the stratum division model is obtained by fusing U-Net and Bi-LSTM. Finally, a sliding window is used to eliminate noise points from an initial label sequence, and the target label sequence is obtained by correction based on a geological monotonicity constraint. The application effectively overcomes the problem of insufficient space-time coordination ability of a single model by extracting multi-scale space features through one-dimensional U-Net and capturing stratum time sequence dependency relationships through Bi-LSTM, and significantly improves the accuracy of stratum boundary identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of stratum identification and division, and more particularly to a well logging stratum division method and device based on hybrid deep learning and geological constraints. BACKGROUND

[0002] Well logging stratum division is a basic work of oil and gas exploration and development, which mainly relies on accurate identification of stratum boundaries by well logging data, and the division results directly affect geological modeling, reservoir analysis and development plan. The traditional stratum division method mainly depends on the manual interpretation of experts on multiple well logging curves, which is highly subjective and low in efficiency, and is difficult to cope with the challenges brought by modern high-density well network and massive data. Although semi-automatic division methods based on rules have emerged to some extent, they have improved efficiency, but they mostly use static heuristic rules, which have limited robustness and applicability in the face of complex and nonlinear stratum changes.

[0003] With the development of deep learning technology, convolutional neural networks (CNN), recurrent neural networks (RNN), long short-term memory networks (LSTM) and U-Net models have been applied to automatic well logging stratum division, significantly improving feature extraction capability and division efficiency. However, these models have the problems of insufficient spatial and temporal feature modeling, discontinuous boundary prediction and insufficient geological constraints when dividing strata. SUMMARY

[0004] Therefore, the present application provides a well logging stratum division method and device based on hybrid deep learning and geological constraints to solve the problems of insufficient spatial and temporal feature modeling, discontinuous boundary prediction and insufficient geological constraints in existing well logging stratum division methods.

[0005] To achieve the above-mentioned purpose, the present scheme is as follows:

[0006] A well logging stratum division method based on hybrid deep learning and geological constraints, comprising:

[0007] preprocessing well logging data to obtain initial data, the well logging data including well logging curves and stratum marker point depths;

[0008] inputting the initial data into a pre-trained stratum division model for prediction to obtain an initial label sequence, the stratum division model being obtained by fusing a U-Net and a Bi-LSTM;

[0009] adopting a sliding window to eliminate noise points from the initial label sequence and correcting it based on a geological monotonicity constraint to obtain a target label sequence.

[0010] Preferably, the process of preprocessing the well logging data includes:

[0011] Calculate the relative depth of each formation marker point based on the reference marker point;

[0012] According to the relative depth of each formation marker point, the formation is label divided;

[0013] The length of the well logging curve of each well is unified by zero padding.

[0014] Preferably, the U-Net comprises an encoder and a decoder;

[0015] The encoder extracts local and global spatial features of the well logging curve through multi-scale convolution;

[0016] The decoder up-samples the encoded features through multi-layer up-sampling and feature reconstruction to obtain a spatial feature map;

[0017] The Bi-LSTM is connected to the bottleneck layer of the deepest part of the encoder of the U-Net, and is used to capture the long-term bidirectional dependency relationship of the spatial feature map.

[0018] Preferably, the process of removing noise points from the initial label sequence by using a sliding window and correcting based on a geological monotonicity constraint comprises:

[0019] Smooth the initial label sequence by a sliding window of a preset size, replace the label of each formation marker point with the label with the highest frequency in the window, so as to remove isolated noise points;

[0020] Based on the geological monotonicity constraint, the formation label sequence is corrected to a corrected label sequence that is monotonically progressive from top to bottom;

[0021] Smooth the corrected label sequence by a sliding window of a preset size to obtain a target label sequence.

[0022] Preferably, the process of correcting the formation label sequence to a corrected label sequence that is monotonically progressive from top to bottom based on the geological monotonicity constraint comprises:

[0023] If the current iteration label is less than the current dominant label, update the current iteration label to the dominant label;

[0024] If the current iteration label is greater than the current dominant label by 2 or more, increase the current dominant label by 1 and limit the label within the target range.

[0025] Preferably, it further comprises:

[0026] According to the target label sequence, extract the predicted depth of each formation marker point, which is the absolute depth corresponding to the first occurrence of the label;

[0027] Record the target label sequence of the predicted well, the predicted depth and the actual depth of each formation marker point, and the absolute error to form a prediction result report.

[0028] The well logging formation division device based on the hybrid deep learning and geological constraint comprises:

[0029] The data processing unit pre-processes the well logging data to obtain initial data, wherein the well logging data comprises well logging curves and formation marker point depths.

[0030] The label prediction unit inputs the initial data into a pre-trained formation division model to obtain an initial label sequence, wherein the formation division model is obtained by fusing a U-Net and a Bi-LSTM.

[0031] The label correction unit eliminates noise points from the initial label sequence by using a sliding window and corrects the initial label sequence based on a geological monotonicity constraint to obtain a target label sequence.

[0032] According to the specific embodiments provided by the present application, the following technical effects are disclosed:

[0033] The well logging formation division method based on the hybrid deep learning and geological constraint provided by the present application pre-processes well logging data, wherein the well logging data comprises well logging curves and formation marker point depths; inputs initial data into a pre-trained formation division model to obtain an initial label sequence, wherein the formation division model is obtained by fusing a U-Net and a Bi-LSTM; and eliminates noise points from the initial label sequence by using a sliding window and corrects the initial label sequence based on a geological monotonicity constraint to obtain a target label sequence. The present application effectively overcomes the problem of insufficient space-time coordination ability of a single model by extracting multi-scale spatial features by a one-dimensional U-Net and capturing a stratigraphic time sequence dependency relationship by a Bi-LSTM, and significantly improves the accuracy of formation boundary identification.

[0034] The well logging formation division method based on the hybrid deep learning and geological constraint provided by the present application eliminates noise points from the initial label sequence by using a sliding window and corrects the initial label sequence based on a geological monotonicity constraint. The present application can effectively improve the accuracy and geological rationality of formation boundary identification by structure optimization and geological rule post-processing constraints. BRIEF DESCRIPTION OF DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, brief descriptions will be given below to the drawings needed to be used in the embodiments or prior art descriptions. Obviously, the drawings in the following description are only embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of the provided drawings.

[0036] Figure 1A flow chart of the well logging formation division method based on the mixed deep learning and geological constraints is provided for the embodiments of the present application.

[0037] Figure 2 A well logging curve cycle statistical chart is provided for the embodiments of the present application.

[0038] Figure 3 A well logging curve data thermal map is provided for the embodiments of the present application.

[0039] Figure 4 A formation division model prediction flow chart is provided for the embodiments of the present application.

[0040] Figure 5 A formation division model training process accuracy chart is provided for the embodiments of the present application.

[0041] Figure 6 A formation division model training process training loss chart is provided for the embodiments of the present application.

[0042] Figure 7 A test well formation comparison chart is provided for the embodiments of the present application.

[0043] Figure 8 A formation boundary prediction error distribution chart is provided for the embodiments of the present application.

[0044] Figure 9 A well logging formation division device schematic diagram based on the mixed deep learning and geological constraints is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0046] First, combined with the geological characteristics of the well logging data, the well logging formation division method based on the mixed deep learning and geological constraints is introduced. Figure 1 The method includes the following steps. Figure 1

[0047] Step S1, pre-processing the well logging data.

[0048] Specifically, the well logging data is collected, and the well logging data includes well logging curves and formation marker point depths. The well logging curves can effectively reflect the formation cycle characteristics. The layered data is loaded to obtain the depth values of the key formation marker points of each well. The well logging curves common to all wells are taken as model input features, and if a well lacks effective well logging curves, it is removed from the well list. As shown in Figure 2 ​As shown, the well logging data of the training wells (W1, W2, W3) and the prediction well (W4) in the study area is collected, and the extracted well logging curve cycle characteristics are analyzed to count the number of strong cycles and weak cycles, reflecting the periodic characteristics of stratum deposition. The layered data is loaded to obtain the depth values of the key stratum marker points (T1, T2, T3) of each well.

[0049] The well logging data of the wellbore is preprocessed by interpolation, normalization, noise suppression, etc., and then the well logging curve data is divided by a sliding window to obtain initial data. The initial data is converted into a sample sequence form to input the neural network model. The label corresponding to each window sample is a stratigraphic boundary point or a stratum category.

[0050] First, the relative depths of each stratum marker point are calculated based on the reference marker points to eliminate the absolute depth deviation caused by structural differences and unify the depth reference standard. For example: taking the T1 marker point of each well as the reference, the relative depths of all stratum marker points are calculated (relative depth = absolute depth - T1 depth).

[0051] Then, according to the relative depths of each stratum marker point, the strata are divided into labels to realize the category annotation of the strata. For example: according to the relative depth range of T1, T2, and T3, the strata are divided into 4 labels (0-3), where the relative depth < 0 is label 0, the interval from 0 to the first relative depth (T2-T1) is label 1, the interval from the first relative depth to the second relative depth (T3-T1) is label 2, and the part greater than the second relative depth is label 3.

[0052] Finally, the well logging curve is standardized to make the data meet the model input requirements, and the length of each well logging curve is unified by zero padding to ensure the consistency of the model input size. For example: the data sequence length of each well can be unified to 4 times the maximum sequence length, and the distribution of the standardized well logging curves after standardization is as shown in Figure 3 which can intuitively observe the change rule of the curve.

[0053] Step S2, input the initial data into the stratigraphic division model pre-trained to generate the stratigraphic category probability sequence of each stratum marker point to obtain the initial label sequence.

[0054] Specifically, the initial data of the well to be divided is input into the trained stratigraphic division model to generate the stratigraphic category probability sequence of each stratum marker point to obtain the initial label sequence. The stratigraphic division model is obtained by fusing U-Net and Bi-LSTM, and the stratigraphic division model structure includes an encoder, a bottleneck layer, and a decoder. As shown in Figure 4 which uses a one-dimensional U-Net network to extract the multi-scale spatial features of the well logging curve, and a Bi-LSTM through the bottleneck layer to capture the time sequence dependence of the stratum sequence.

[0055] The encoder part of the one-dimensional U-Net network adopts a plurality of one-dimensional convolution (Conv1D) and pooling (MaxPooling1D) layers, extracts local and global spatial features of the well logging curve through multi-scale convolution operation. The decoder part adopts a plurality of up-sampling (UpSampling1D) and one-dimensional convolution layers, and fuses the feature maps of the corresponding levels of the encoder to realize accurate positioning of the formation boundary. The structure is as follows:

[0056]

[0057] wherein, is the input well logging curve sequence, the encoder extracts features and down-samples the input sequence to obtain encoded features, is the decoder up-samples and reconstructs the encoded features, is the output spatial feature map.

[0058] The bottleneck layer is arranged at the deepest part of the encoder, adopts a bidirectional long short-term memory network (Bi-LSTM) layer, is used to capture and learn the long-term bidirectional dependency relationship of the spatial feature map, and is used to enhance the modeling ability of the stratum context evolution relationship. The prediction output is the stratum label or boundary probability of each stratum marker point:

[0059] wherein, is the output spatial feature map, is the bidirectional long short-term memory network processing feature sequence, represents converting the output into a probability distribution, is a probability matrix (size ) representing the probability of each stratum marker point belonging to the C class of strata, taking the class corresponding to the maximum probability of each stratum marker point as the label to obtain an initial label sequence.

[0060] Step S3, a sliding window is used to eliminate noise points from the initial label sequence, and is modified based on the geological monotonicity constraint.

[0061] Specifically, as Figure 4 shown, the geological rules are applied for post-processing correction, a sliding window mode filter is used to eliminate noise points, and the stratum label sequence is forced to meet the self-top-down monotonicity rule to identify the accurate stratum boundary depth.

[0062] Firstly, the initial label sequence is smoothed by a sliding window of a predetermined size, and the label of each stratum marker point is replaced by the label with the highest frequency in the window to eliminate isolated noise points. That is, based on the class label sequence with the maximum probability, a moving window mode filter is used for preliminary smoothing:

[0063]

[0064] wherein, is the original predicted label of the i-th stratigraphic marker point, is the i-th stratigraphic marker point, is the width of the sliding window, is the label sequence in the window centered at the i-th stratigraphic marker point with a width of is the label sequence in the window centered at the i-th stratigraphic marker point with a width of is the label sequence in the window centered at the i-th stratigraphic marker point with a width of is the class with the highest frequency of occurrence in the window.

[0065] Then, based on the geological monotonicity constraint, the stratigraphic label sequence is corrected to a self-top-down monotonically progressive corrected label sequence. The smoothed label sequence is traversed from top to bottom, and if the current traversal label is less than the current dominant label (violating the deposition order), the current traversal label is updated to the dominant label, and if the current traversal label is greater than the current dominant label by 2 or more, the current dominant label is increased by 1, and the label is limited within the target range to avoid outliers.

[0066] If the stratigraphic label of the current stratigraphic marker point is less than the stratigraphic label of the previous current stratigraphic marker point, the current point label is forced to be corrected to be the same as the previous depth point to ensure that the predicted stratigraphic sequence strictly follows the monotonicity rule of geological deposition. According to the stratigraphic self-top-down progressive rule, a constraint function is introduced to force adjustment of the reverse jump sequence, so that the final sequence satisfies: .

[0067] wherein, is the stratigraphic class label of the i-th depth point from shallow to deep (1st to i-th). r

[0068] Finally, the corrected label sequence is smoothed by a sliding window of a predetermined size to obtain a target label sequence. The corrected label sequence is again filtered by a sliding window mode to further improve the continuity and stability of the sequence, and the final stratigraphic prediction label is obtained.

[0069] The logging stratigraphic division method based on the hybrid deep neural structure preserves the local structure expression capability of U-Net, introduces a bidirectional time modeling mechanism and a reasonable post-processing correction strategy, and realizes fine identification and intelligent division of stratigraphic boundaries. Compared with traditional single structure models based on CNN or RNN, the method has higher precision, robustness and geological rationality in actual multi-well applications.

[0070] Further, in order to better analyze the geological division result, the logging stratigraphic division method of the embodiment of the present application can further perform the following steps based on the foregoing embodiment:

[0071] ​According to the target label sequence, the predicted depth of each stratigraphic marker point (T1, T2, T3) is extracted, that is, the absolute depth corresponding to the first occurrence of the label. After completing the post-processing of the geological rules, the final boundary depth between two adjacent strata is determined by searching for the first depth point at which the first upward jump of the label value occurs in the corrected stratigraphic label sequence.

[0072] The target label sequence of the predicted well, the predicted depth and the actual depth of each stratigraphic marker point, and the absolute error are recorded to form a prediction result report.

[0073] The division result output by the embodiment of the present application can generate a structured stratigraphic distribution table, record the start and end depth, thickness and category label of each stratum, and can be compared with the existing artificial horizon to evaluate the accuracy, Kappa coefficient, average horizon deviation and other performance indicators.

[0074] Next, the training process of the stratigraphic division model is introduced, as shown in Figure 4 .

[0075] The preprocessed training well data with stratigraphic labels are taken as input, and the stratigraphic labels are converted into one-hot format; the classification cross-entropy is used as the loss function, and the model parameters are optimized and updated through the back propagation algorithm.

[0076] ① Model construction:

[0077] A stratigraphic division model based on U-Net is constructed, which includes an encoder, a bottleneck layer and a decoder.

[0078] The encoder is composed of two groups of convolution modules, each group containing two layers of one-dimensional convolution (convolution kernel size 3, activation function ReLU, He normal distribution initialization weight, padding mode same) and one layer of maximum pooling (step 2), which gradually extracts the local features of the well logging data (such as lithology sudden change points).

[0079] The bottleneck layer adopts bidirectional LSTM (hidden unit number 128, return sequence True), which can capture the bidirectional time sequence dependence relationship of the stratigraphic sequence (such as sedimentary cycle rhythm).

[0080] The decoder includes two groups of up-sampling and convolution modules, which recover the feature map size through up-sampling (step 2) and fuse with the corresponding layer features of the encoder, realize the complementation of high and low layer features, and finally output the probability distribution of four types of strata through one layer of convolution (kernel size 1, activation function softmax).

[0081] ② Model parameter configuration:

[0082] The input shape is the length of the filled sequence;

[0083] The Adam optimizer (learning rate 1e-4) is used, the classification cross-entropy is used as the loss function, the loss function is categorical_crossentropy, the accuracy is used as the evaluation index, and the model parameters are optimized and updated through the back propagation algorithm.

[0084] ③Model training:

[0085] The training well data with the formation label after preprocessing is taken as the training set, and the formation label is converted into one-hot format; the prediction well data is taken as the validation set to input the model, the training round is 100, and the batch size is 1.

[0086] The accuracy and loss of training and validation are monitored in real time during the training process, the model convergence is shown through the smoothed curve, and the round in which the peak value of the validation accuracy and the valley value of the validation loss are marked, and the generalization ability of the model is evaluated. Figure 5 The formation division model training process accuracy graph of the embodiment of the application; Figure 6 The formation division model training process training loss graph of the embodiment of the application. Based on the curve, the training state of the model can be intuitively reflected, and the marking of the peak value of the validation accuracy and the valley value of the validation loss helps to determine the best training effect of the model.

[0087] Further, the feasibility of the logging formation division method based on mixed deep learning and geological constraints of the embodiment of the application is verified.

[0088] As shown in Figure 7 , the main view and the local enlarged view clearly show the comparison of the actual and predicted formation boundaries, and the prediction deviation of each marked point is within 0.5m.

[0089] As shown in Figure 8 , the prediction absolute errors of T1, T2 and T3 are mainly concentrated in the 0-0.3m interval, of which the 0-0.2m interval accounts for 65%, the 0.2-0.3m interval accounts for 25%, and the 0.3-0.5m interval accounts for 10%, reflecting that the logging formation division method of the embodiment of the application has high prediction accuracy.

[0090] The logging formation division device based on mixed deep learning and geological constraints provided by the embodiment of the application is described below, and the logging formation division device based on mixed deep learning and geological constraints described below can be correspondingly referred to the logging formation division method based on mixed deep learning and geological constraints described above.

[0091] Firstly, the logging formation division device based on mixed deep learning and geological constraints is introduced in combination with Figure 9 , as shown in Figure 9 , the logging formation division device based on mixed deep learning and geological constraints can comprise:

[0092] The data processing unit 100 pre-processes the logging data to obtain initial data, the logging data including logging curves and formation marker point depths;

[0093] The label prediction unit 200 inputs the initial data into a pre-trained formation division model to obtain an initial label sequence, the formation division model being obtained by fusing a U-Net and a Bi-LSTM;

[0094] The label correction unit 300 removes noise points from the initial label sequence by using a sliding window and corrects the initial label sequence based on a geological monotonicity constraint to obtain a target label sequence.

[0095] Finally, it should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one from another entity or action without necessarily requiring or implying any actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises... a" does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0096] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts between the various embodiments can be mutually referred to.

[0097] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for well-logging stratigraphic division based on hybrid deep learning and geological constraints, characterized in that, The method comprises the following steps: preprocessing well logging data to obtain initial data, wherein the well logging data comprises well logging curves and formation marker point depths; inputting the initial data into a pre-trained formation division model to obtain an initial label sequence, wherein the formation division model is obtained by fusing a U-Net and a Bi-LSTM; eliminating noise points in the initial label sequence by using a sliding window and correcting the label sequence based on a geological monotonicity constraint to obtain a target label sequence; the U-Net comprises an encoder and a decoder; the encoder extracts local and global spatial features of the well logging curves by using multi-scale convolution; the decoder performs upsampling and feature reconstruction on the encoded features by using multi-layer upsampling to obtain a spatial feature map; the Bi-LSTM is connected to a bottleneck layer at the deepest part of the encoder of the U-Net and is used to capture long-term bidirectional dependency relationships of the spatial feature map; the process of eliminating noise points in the initial label sequence by using a sliding window and correcting the label sequence based on a geological monotonicity constraint comprises the following steps: smooth the initial label sequence by using a sliding window of a preset size, and replace the labels of the formation marker points with the label that appears most frequently in the window to eliminate isolated noise points; correct the label sequence based on the geological monotonicity constraint to obtain a corrected label sequence that is monotonically progressive from top to bottom; smooth the corrected label sequence by using a sliding window of a preset size to obtain the target label sequence; the process of correcting the label sequence based on the geological monotonicity constraint to obtain a corrected label sequence that is monotonically progressive from top to bottom comprises the following steps: traverse the smoothed label sequence, and if the current traversed label is less than the current dominant label, update the current traversed label to the dominant label; if the current traversed label is greater than the current dominant label by 2 or more, increase the current dominant label by 1 and limit the label within a target range.

2. The method of claim 1, wherein the method is a hybrid deep learning and geologically constrained well stratigraphic division method. the process of preprocessing the well logging data comprises the following steps: calculate the relative depths of the formation marker points based on the reference marker points; label divide the formations according to the relative depths of the formation marker points; unify the lengths of the well logging curves of the wells by using zero padding.

3. The method of claim 1 or 2, wherein the method further comprises: The method further comprises the following steps: extract the predicted depths of the formation marker points, i.e., the absolute depths corresponding to the first occurrence of the labels, according to the target label sequence; record the target label sequence of the predicted well, the predicted depths and actual depths of the formation marker points, and the absolute error to form a prediction result report.

4. A logging formation division device based on hybrid deep learning and geological constraints, characterized in that, The method comprises the following steps: a data processing unit preprocesses well logging data to obtain initial data, wherein the well logging data comprises well logging curves and formation marker point depths; a label prediction unit inputs the initial data into a pre-trained formation division model to obtain an initial label sequence, wherein the formation division model is obtained by fusing a U-Net and a Bi-LSTM; a label correction unit eliminates noise points in the initial label sequence by using a sliding window and corrects the label sequence based on a geological monotonicity constraint to obtain a target label sequence; the U-Net comprises an encoder and a decoder; the encoder extracts local and global spatial features of the well logging curves by using multi-scale convolution; The decoder up-samples the encoded features and reconstructs features by multi-layer up-sampling to obtain a spatial feature map; The Bi-LSTM is connected with a bottleneck layer at the deepest part of the encoder of the U-Net, and is used to capture long-term bidirectional dependency of the spatial feature map; The process of removing noise points from the initial label sequence by using a sliding window and correcting based on a geological monotonicity constraint comprises: Smooth the initial label sequence by using a sliding window of a preset size, replace the label of each stratigraphic marker point with the label with the highest frequency in the window, to remove isolated noise points; Correct the stratigraphic label sequence to a corrected label sequence that is monotonically progressive from top to bottom based on a geological monotonicity constraint; Smooth the corrected label sequence by using a sliding window of a preset size to obtain a target label sequence; The process of correcting the stratigraphic label sequence to a corrected label sequence that is monotonically progressive from top to bottom based on a geological monotonicity constraint comprises: Traverse the smoothed label sequence, if the current traversal label is less than the current dominant label, update the current traversal label to the dominant label; If the current traversal label is greater than the current dominant label by 2 or more, add 1 to the current dominant label, and limit the label within a target range.

Images